Ethics and Trusted AI
Ethical Aspects of AI Implementation
In 2021, HSE University became a signatory to the Code of Ethics in the field of AI. Its most important feature is a human-centred and risk-oriented approach to understanding the prospects for the development of artificial intelligence. In 2024, HSE University developed and approved the Declaration of Ethical Principles for the Use of AI. The work of the AI Centre also aligns with the principles of this declaration, particularly in that the development and use of AI must be safe, meaning they should ensure the protection of personal data, help prevent errors, and minimise risks.
The AI Centre adheres to ethical principles in the development and implementation of artificial intelligence systems:
Principle of Safety
Protection of personal data, prevention of errors, and minimisation of risks
Principle of Transparency and Fairness
Ensuring the comprehensibility of algorithms and predictability of their decisions
Principle of Preventing Discrimination Against Various Social Groups
Periodic updating of data sets to reduce the risk of bias
Principle of Control and Accountability
Monitoring the quality of AI performance and identifying potential issues
The implementation of solutions developed based on large language models potentially brings a number of risks that should be addressed in a timely manner. To minimise these risks, the AI Centre has developed a comprehensive set of prevention and response measures:
Risks
- Risk of uncertainty due to potential AI errors
- Risk of potential bias in assessments due to the shortcomings of training datasets or data used for augmenting large language models without appropriate adaptation (RAG approach)
Solutions
- Ensuring the monitoring of the quality of AI technologies, including LLMs, to identify possible model hallucinations and prevent the occurrence of errors/data leaks
- Applying a structured approach to the periodic updating of text data sets to combat potential bias in AI tools used for decision-making
Ensuring the Trustworthy Nature of AI Technologies
Research in the development of tailored industry-specific AI models will focus on ensuring the reliability of the information generated through their use, including fact-checking and combating hallucinations, particularly when working with data in the Russian language.
The development of the solution will include a results verification system, as well as the creation of benchmarks in accordance with international standards. The evaluation of models will be conducted using widely accepted metrics, which will allow for the selection of models demonstrating the best performance. Metrics will be adapted for specific tasks and application contexts.
At the AI Centre, secure artificial intelligence systems are being developed:
A software solution for the automatic detection and/or neutralisation of attacks on data used in training AI models (anomaly detection, data poisoning, unbalanced samples, attacks on data distribution, outliers, hidden patterns)
Datasets and the selection of AI models for conducting trials and testing the developed methodology
Software for the protection of various classes of AI models, including protection against model stealing attacks, and a recommendation model for mitigating threats
We strive to create safe, transparent, and responsible artificial intelligence that meets high ethical standards and is aimed at increasing trust among users and society as a whole.